Using elementary disturbances for testing of machine learning models A general method for testing of machine learning models based on elementary disturbances: An evaluation with image and audio data

Abstract

This thesis explores the testing of machine learning models. The problem with current testing methods is that testing often is case-specific and require significant additional effort to perform. A novel method of adding simple elementary disturbances to the input data is used. The method is done in a general way that should work for different kinds of data and different types of machine learning models. The simple disturbances can be used to predict how well a machine learning model handles unseen disturbances. A general testing methodology could be useful as a simple prediction of a machine learning model’s resilience to unseen disturbances.